The Evolving Relationship Between Ethics and Safety in AI
As AI systems support more distant training supervision, wider action space ranges, more powerful optimization, more general multi-task algorithms, and more meta specification, the dependencies between AI ethics and AI safety become more pronounced. We discuss why architectural foresight for both safety and ethical concerns is both possible and warranted, and how technical architecture, professional ethics, safety engineering, and stakeholder interests are increasingly inextricably bound together. Through the paradigm of technical value alignment we explore ways to predict new classes of misalignment bugs and ways to prevent or mitigate them as early as possible.
Richard Mallah is Director of AI Projects at Future of Life Institute, an existential risk mitigation and technology beneficence NGO, where he works on the robust, safe, beneficent development of advanced AI. He helps move the world toward existential hope and away from outsized risks via meta-research, analysis, research organization, community building, and advocacy, with respect to technical progress, strategy, and policy coordination. Richard is also part of IEEE’s initiative on autonomous systems ethics, a committee member at Partnership on AI, a senior advisor to the The Future Society, and head of AI R&D at startup Avrio AI.